SmartPhotoCrafter: Unified Reasoning, Generation and Optimization for Automatic Photographic Image Editing
Ying Zeng, Miaosen Luo, Guangyuan Li, Yang Yang, Ruiyang Fan, Linxiao Shi, Qirui Yang, Jian Zhang, Chengcheng Liu, Siming Zheng, Jinwei Chen, Bo Li, Peng-Tao Jiang
cs.CV
2026-04-21
Editing models need you to tell them what to change. SmartPhotoCrafter uses an Image Critic to reason about a photo's flaws (haze, exposure, saturation) and suggest fixes, then a Photographic Artist edits accordingly. Across three training stages (pretrain, reasoning adaptation, coordinated RL) it beats a field of generative baselines on automatic enhancement and multi-edit instruction adherence.
Traditional image editing assumes you can instruct it: you must know what is wrong and how to fix it. Non-expert users often cannot give an optimal editing instruction, and instruction-conditioned models have no ability to perceive image quality or diagnose deficiencies, so fully automatic enhancement is out of reach. This paper makes the model see what is wrong with a photo and fix it itself.
SmartPhotoCrafter casts editing as a tightly coupled reasoning-to-generation process with two modules.
The Image Critic, a multimodal language model, comprehends image quality, identifies deficiencies (haze, desaturation, low contrast), scores quality, and emits edit suggestions (reasoning in a think block first). The Photographic Artist consumes the Critic's latent reasoning representations (a representation-level coupling, not just passed text) to produce targeted, faithful edits.
Training is three stages: Stage I foundation pretraining with SFT on image-quality-assessment and editing data; Stage II reasoning-conditioned adaptation where the Artist is conditioned on the Critic's latents; and Stage III coordinated reasoning-to-generation RL that jointly trains both modules. A Photometric Control Reward (rphoto) targets attribute-level tonal and color sensitivity to stop RL from drifting toward over-optimization. The pipeline emphasizes photo-realistic generation while supporting restoration and retouching with consistent color and tone semantics.
Data is a custom staged dataset: the Critic uses KonIQ-10K, SPAQ, and KADID-10K plus distortion data (deblurring, dehazing, low-light, shadow removal), and retouching uses FilmSet for synthetic color grading and RealBokeh and BokehDiff for depth of field.
Against Instruct-Pix2Pix, FLUX2.Dev, Qwen-Image-Edit, OmniGen2, and Step1X-Edit (reference GT: MUSIQ 70.96, NIMA 5.23).
| Method | MUSIQ | FID | LPIPS | DINO | CLIP |
| FLUX2.Dev | 72.94 | 76.26 | 0.26 | 0.93 | 0.88 |
| Qwen-Image-Edit | 68.63 | 42.81 | 0.17 | 0.96 | 0.93 |
| Ours | 69.52 | 27.96 | 0.10 | 0.98 | 0.96 |
The method ranks second on perceptual MUSIQ behind FLUX2, but FLUX2 looks AI-stylized with high FID; the method is best on distributional consistency (lowest FID and LPIPS) and semantic alignment (highest DINO and CLIP), and is more balanced. On multi-edit instruction adherence it is first on all six metrics (PSNR 21.05, SSIM 0.82, LPIPS 0.09), and on deblurring and dehazing it is mostly best. The ablation confirms rphoto: adding RL without it improves MUSIQ slightly but worsens FID from 30.61 to 38.51 (distribution drift), while the full reward brings FID back to 27.96 and lifts every other metric.
Instruction-free, self-diagnosing photo enhancement is a real pain point, especially for ordinary phone users. This couples multimodal reasoning and generation at the representation level rather than just feeding the generator the Critic's text, and it uses a photometric reward to restrain the over-enhancement RL tends to drift into. The diagnose-then-edit coupling and reward design are worth borrowing for image-editing and computational-photography products.
There is no dedicated limitations section, and a few gaps stand out. Evaluation is entirely on proxy metrics (MUSIQ, NIMA, PSNR, FID, LPIPS, CLIP, DINO) with no human-preference study, while editing is inherently subjective with many valid outputs. The method is strong on fidelity and distributional consistency at the cost of not topping perceptual-quality scores, so it may not suit highly stylized or heavily reworked edits. No base architecture details or comparison with the strongest closed editing models are given; only open generative baselines are compared, and the automatic-enhancement test set includes synthetic degradations, so robustness on real messy phone photos is unknown.